You don't need error nodes, you need better features

According to LessWrong, researcher Evan Lloyd has published work on AI interpretability with a new approach to a persistent challenge. Sparse autoencoders break down a model's computations into interpretable features, but stacking them across multiple layers causes errors to multiply so severely the model stops functioning. Lloyd's solution, replacement-aware training, teaches autoencoders to tolerate upstream errors rather than requiring corrections afterward. On Gemma-2-2B, this method maintains model capability and produces valid responses where standard techniques fail. Code and trained weights are now publicly available.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton